NNP: a neural net classifier using prototypes
Christine Decaestecker · 2002
A three-layer neural net classifier for multiclass object recognition problems requiring piecewise nonlinear discriminant surfaces is presented. The hidden layer is composed of prototypes of each class. The weights from the input to the hidden layer are the vector descriptions of prototypes (in the input feature space). The output layer neurons represent the classes, the hidden-to-output weights being binary and fixed. They map each prototype neuron to one of the class output neurons. Only the input-to-hidden weights are adapted by an algorithm using deterministic annealing and gradient descent techniques. This algorithm permits the distribution of prototypes in classes while minimizing the classification error rate.>